I Started Using an AI Closet Before Buying Clothes — Here’s What It Can and Can’t Tell You

gxlbfc — Mon, 29 Jun 2026 08:45:00 +0000

I used to shop in one of two ways:

I saw an outfit on someone else and tried to recreate it.
I found one item I liked and convinced myself I would somehow build outfits around it later.

Both methods produced the same result: a wardrobe full of individually nice pieces that did not always work together — or feel like me.

So I added a step before buying anything:

I test the style direction with AI first.

Not the exact size. Not the fabric quality. Not whether a particular pair of trousers will pinch at the waist.

I use AI virtual try-on to answer a simpler question:

Is this style worth exploring on me before I spend money on it?

That distinction matters, because AI can be surprisingly useful as a visual filter — and very misleading if you treat it like a fitting room.

The shift: test a direction, not a product

The biggest improvement came when I stopped asking:

Would this exact jacket look good on me?

and started asking:

Do I actually like myself in this silhouette, palette, and level of formality?

The first question is difficult for AI. It requires accurate sizing, construction, material behavior, and product fidelity.

The second question is much more realistic. It is about visual direction.

For example, an “old money” look is not just one camel coat. It is a combination of:

restrained neutral colors
clean tailoring
longer, quieter silhouettes
low-contrast accessories
polished rather than trend-heavy styling

If I like that complete direction on myself, I can shop more intelligently. I may not buy the exact AI-generated outfit, but I now know that camel, cream, soft knitwear, straight tailoring, and brown leather are useful signals.

That is already more valuable than adding another random “nice top” to a cart.

My 10-minute pre-shopping test

Here is the workflow I use.

1. Start with one neutral photo

I use the same clear, front-facing photo for every test:

natural standing pose
arms not covering the torso
even lighting
most of the body visible
no oversized coat hiding the original silhouette

Keeping the person photo fixed is important. If the pose, camera angle, and lighting change every time, I end up comparing photographs instead of comparing clothes.

2. Test several clearly different style lanes

I do not begin with ten versions of almost the same outfit. I choose directions that are far enough apart to teach me something.

For example:

minimal casual
denim casual
polished office wear
old money
date night

The goal is not to crown one permanent personal aesthetic. Most people need more than one.

The useful question is: Which directions feel natural, and which feel like a costume?

3. Look for repeated signals

One AI image means very little. Repeated preferences are more useful.

After several tests, I write down what keeps working:

Do I prefer a defined waist or a straighter silhouette?
Do warm neutrals suit the overall impression better than cool grey?
Do I like sharp shoulders or softer layers?
Do ankle-length trousers feel better than wide, floor-length shapes?
Do high necklines make the look feel refined or restrictive?
Do I consistently prefer low-contrast outfits?

These observations turn a vague reaction — “I like it” — into a shopping filter.

Instead of searching for “cute work clothes,” I can look for:

warm-neutral blazer, softly structured shoulder, single-breasted, hip length

That is a much better starting point.

4. Separate admiration from identification

This was the most useful lesson.

There are outfits I love looking at but do not want to wear.

AI makes that difference obvious because it moves the outfit from a model, mood board, or product page onto something closer to my own visual context.

Sometimes the reaction is immediate:

“Beautiful, but not me.”
“I like the color, not the silhouette.”
“This works only because of the styling.”
“I would wear this if the jacket were shorter.”

That is not a failed result. That is the result.

What AI can tell you

Used carefully, AI virtual try-on can help with four things.

1. Whether a color direction is worth testing in real life

It can give a rough sense of whether an outfit feels harmonious, heavy, washed out, too severe, or surprisingly balanced.

This is not professional color analysis. Screens, lighting, image processing, and the model itself can all shift colors.

But it can tell me whether “more warm beige” is a promising direction before I order four beige items.

2. Whether a silhouette feels like you

AI is useful for comparing broad shapes:

cropped versus long outerwear
fitted versus relaxed
structured versus draped
high-waisted versus low-rise
minimal versus heavily layered

Again, it is directional. It does not know the actual garment measurements or how the fabric behaves on your body.

3. Whether a style works with your existing identity

Clothing is not only about body shape. Hair, posture, personal energy, work environment, and how formal you like to feel all matter.

Seeing yourself near a style can expose whether you enjoy wearing it or only enjoy the idea of it.

4. What is missing from your wardrobe

If several successful looks depend on the same type of item — perhaps a cream blazer, straight-leg denim, or a simple dark evening dress — that repeated item may be a genuine wardrobe gap.

That is much more actionable than buying whatever happens to be trending.

What AI cannot tell you

This is the part every virtual try-on article should include.

1. Your correct size

A generated image is not a measurement tool. It cannot reliably tell you whether a real garment will button, pull, gap, or need tailoring.

Always use the retailer’s measurements, garment dimensions, reviews, and return policy.

2. Fabric quality

AI can render beautiful wool, silk, denim, or linen. That says nothing about the item arriving at your door.

It cannot detect thin lining, scratchy knitwear, weak seams, cheap hardware, or fabric that becomes transparent in daylight.

3. Real drape and movement

A still image cannot tell you what happens when you sit, walk, lift your arms, or wear the garment for six hours.

This is especially important for fitted dresses, trousers, occasion wear, and anything made from stiff or very light fabric.

4. Exact color

Product photography is already affected by lighting and editing. AI adds another interpretation layer.

Treat color as a family — warm cream, muted blue, deep burgundy — rather than an exact match.

5. Whether the real item is well made

AI can help decide whether a design direction makes sense. It cannot inspect construction.

That still requires product details, close-up photos, material composition, customer reviews, and sometimes seeing the item in person.

The buying checklist I use now

Before purchasing, I separate the decision into two passes.

Pass 1: visual direction

Does the color family work?
Does the overall silhouette feel natural on me?
Can I name at least three occasions where I would wear it?
Does it work with pieces I already own?
Am I attracted to the outfit, or only to the model and photography?

AI can help here.

Pass 2: physical reality

Are the garment measurements right?
Is the fabric appropriate for the season and use?
Do reviews mention shrinking, pilling, transparency, or poor construction?
Can I move comfortably in this cut?
Is the return policy reasonable?

AI cannot answer these questions.

I only buy when both passes make sense.

Where the AI closet fits

I built Dressora around this style-first workflow.

Instead of requiring people to find and upload a clothing image for every experiment, the AI Closet includes ready-made directions such as minimal casual, office wear, old money, Y2K, denim, date-night looks, wedding outfits, Hanfu, streetwear, and fantasy styles.

You can start broad, notice what works, and then narrow down.

For a specific garment you found elsewhere, you can switch to the AI clothes changer and upload your own outfit reference.

The important part is not generating the most impressive picture.

It is reducing uncertainty before spending money.

Final thought

AI virtual try-on did not replace shopping for me. It inserted a useful pause before shopping.

It helps me move from:

“That outfit looks great.”

to:

“I like the warm palette and long outer layer, but I need a straighter skirt and a less formal shoe.”

That is a better question to take into a store, a product page, or a tailor.

Use AI to explore style, eliminate weak ideas, and build a more specific shopping list.

Then use measurements, materials, reviews, and real-world try-on to make the final decision.

That is the boundary where AI becomes genuinely helpful — without pretending it knows more than it does.

Building an AI Clothes Changer: provider abstraction, async jobs, and a credit system that won't lose money

gxlbfc — Wed, 17 Jun 2026 08:16:53 +0000

I recently launched Dressora, an AI clothes changer that swaps outfits onto a single photo for virtual try-on. The product side is fun, but the parts I actually sweated over were the boring backend bits: orchestrating multiple AI providers, handling long-running generation jobs, and building a credit system that never double-charges or loses money. Here's what I learned.

Stack

Next.js 15 (App Router) + React 19 + TypeScript
PostgreSQL + Drizzle ORM
Cloudflare R2 for media storage
Multiple AI image/video providers behind one interface

1. Don't marry a single AI provider

AI providers change pricing, rate limits, and quality constantly. Hardcoding one is a trap. I put everything behind a small factory:

const provider = getProvider("evolink");
const task = await provider.createTask({ prompt, aspectRatio });

Each provider implements the same interface (createTask, handleCallback, status mapping). Swapping or adding a provider is a new file, not a refactor. When one provider had an outage, switching the default was a one-line env change.

2. Generation is async — embrace callbacks

AI generation takes 10s–minutes. Blocking a request is a non-starter. The flow:

generate() — create a DB record, freeze credits, call the provider with a callback URL
Provider processes and hits my webhook when done
handleCallback() — download the result, re-upload to R2, mark complete, settle credits

The frontend just polls a lightweight status endpoint. The webhook is the source of truth.

A gotcha: always re-upload the provider's output to your own storage. Provider URLs expire. Downloading and pushing to R2 on completion saved me from dead links later.

3. The credit system was the hardest part

Money + concurrency + async failures = the scariest combination. The pattern that worked: freeze → settle / release.

On request: freeze(credits) — move credits to a "held" state
On success: settle() — actually consume them
On failure/timeout: release() — give them back

freeze  -> hold created, balance reserved
settle  -> hold consumed (success)
release -> hold returned (failure)

This way a failed generation never costs the user, and a user can't fire 10 concurrent jobs with credits for one. I also did FIFO consumption across credit packages so credits with the nearest expiry get used first — fairer for users and simpler for accounting.

4. Lessons

Put external dependencies behind interfaces before you think you need to.
For async jobs, design the failure path first (release credits, retry, timeout) — the happy path is easy.
Re-host anything an external API generates.
A "frozen" intermediate state for credits/money is worth the extra table.

If you want to see the end result, it's live at aiclotheschanger.me. Happy to answer questions about the architecture in the comments.

DEV Community: gxlbfc